Skip to content

v0.18.2

Choose a tag to compare

@github-actions github-actions released this 21 Oct 00:06
· 42 commits to main since this release

Fixed

  • ir update no longer records a file as indexed without its search text. When a collection's preprocessor could not start, the update stored the file's hash but wrote nothing to the full-text index. The file was then not searchable, and every later update reported 0 added, 0 updated, 0 deactivated until --force. Now a preprocessor that cannot start, or stops mid-run, prints a warning and the document is indexed with its raw text, so it stays searchable. The next ir update with a working preprocessor stores those documents again (counted as updated); their vectors are kept. Indexes already damaged this way repair themselves the same way.
  • Concurrent updates of one collection apply each change exactly once. Two updates at the same moment (several ir mcp servers, parallel tool calls in one server, or CLI runs) could fail with UNIQUE constraint failed: documents.path, or with no such table: documents on a new collection.
  • A very long line no longer hangs ir update for collections whose preprocessor streams its output (lines above the OS pipe buffer size).
  • config.yml edits are no longer lost. Commands that change the config (collection add|rm|rename|set-path, preprocessor add|bind|unbind|remove|install) now run one at a time across processes, and the file is replaced atomically, so the daemon and MCP servers never read a half-written file.
  • ir daemon stop never signals a stale pid. After a crash or reboot the pid file could name an unrelated process, which then received SIGTERM. Stop now signals only when the daemon socket answers; otherwise it removes the stale files.
  • Search skips a collection that was never indexed instead of failing the whole query with unable to open database file.
  • ir collection rename moves the index to the new name. Before, the renamed collection was empty and the old index was orphaned.
  • ir collection rm --purge removes every index file (WAL, SHM and the ANN sidecar), not only the .sqlite file.
  • ir preprocessor install re-indexes bound collections with the new preprocessor. It re-indexed before saving the binding, so the first re-index ran without it.
  • ir preprocessor install succeeds without a terminal. When collections existed, a script or agent run failed with prompt: IO error: not a terminal after the install was saved. It now skips the bind prompt and prints the ir preprocessor bind command to use.
  • Per-collection retrieval: settings now apply when ir sync / ir embed build the doc graph and ANN index. They were ignored at build time (only env and defaults were used).
  • The daemon drops a client that connects and then stalls for 10 seconds, so one stuck client cannot block every other search.
  • A slow preprocessor no longer locks other writers out of the collection. ir update preprocessed documents while it held the index write lock, so a second update, ir embed, or the daemon failed with database is locked after 5 seconds. Preprocessing now runs before the lock is taken.
  • A search no longer hangs on a daemon that does not answer. After 120 seconds the search returns the results it already has.
  • Reranking no longer stalls while a collection is being indexed. Saving rerank scores to the cache waited up to 5 seconds for the index write lock; the scores are now skipped when the index is locked.
  • The ANN index is rebuilt after the stored vectors are dropped (ir update --force, or a model with another vector size). Before, vectors that were dropped and embedded again with ANN turned off could leave a sidecar that matched by count only.

Changed

  • The daemon is replaced automatically after an upgrade. A daemon started by the previous binary kept serving until ir daemon stop. Now the next search (or ir daemon start) notices that the daemon's executable was replaced or removed, or that the daemon runs an older version, stops that daemon and starts a new one. That query is answered like a cold start. ir daemon status marks such a daemon. An older or different ir binary never replaces the daemon, so two installs sharing one config dir do not restart it in turns. A long-running ir mcp started before the upgrade keeps working: it starts the daemon from the upgraded binary, and it no longer leaves exited daemons behind as zombie processes.

  • ir search --mode bm25 no longer starts the daemon. It needs no model, so it does not load one in the background.

  • ir sync and ir embed wait when another process is embedding the same collection. The second run then finds nothing pending instead of loading the model and embedding the same documents again.

  • ir update output now explains an empty result. The header line includes the collection path, the counts line starts with the number of files the scan matched, and a third line reports documents that still have no vectors:

    updating 'notes' (/Users/me/notes)…
      321 scanned: 1 added, 0 updated, 0 deactivated
      1 pending embedding (run `ir embed` or `ir sync`)
    

    0 scanned means the path or globs matched no files. ir sync and ir embed print the same header and counts lines (without the pending line, because they embed next). This output is meant to be read, not parsed; a script that matched the old lines exactly needs updating.

  • MCP update result has four new fields per collection: path, scanned, pending_embeddings, and notes. Existing fields are unchanged.

  • MCP callers now get the warnings the CLI prints. update returns them in notes. search adds a second text block, notes:, after the results when the answer was degraded (a collection not indexed, BM25 only while models load, a preprocessor that could not start). The first block is unchanged. Each warning now also says why: a preprocessor that cannot start names the command and the error.

  • MCP update tool description now states its limits: it is text-only (no embeddings), and force: true drops the collection's existing vectors until ir embed or ir sync runs. This matches the existing behavior of ir update --force; no behavior changed.

Dev / Tooling

  • New workflow "CI - evaluation" (.github/workflows/ci-evaluation.yml): runs the model, daemon and network tests and the retrieval-quality checks of scripts/preship.sh on CPU runners (Linux arm64, macOS), nightly and on demand. It is separate from the release workflow.
  • Workflows renamed: "CI" is now "CI - check" and "Release" is now "CD - release". File paths are unchanged.
  • Release candidates: a tag with a - (e.g. v0.19.0-rc.1) creates a GitHub prerelease with the artifacts only. It does not update the Homebrew tap or publish to crates.io.
  • The release check fails when the tag does not match the Cargo.toml version.